io.improbable.keanu.algorithms.mcmc

Class SimulatedAnnealing

java.lang.Object

io.improbable.keanu.algorithms.mcmc.SimulatedAnnealing

public class SimulatedAnnealing
extends java.lang.Object

Simulated Annealing is a modified version of Metropolis Hastings that causes the MCMC random walk to tend towards the Maximum A Posteriori (MAP)

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static interface	SimulatedAnnealing.AnnealingSchedule An annealing schedule determines how T (temperature) changes as a function of the current iteration number (i.e.

Constructor Summary

Constructors

Constructor and Description
SimulatedAnnealing()

Method Summary

Modifier and Type	Method and Description
static SimulatedAnnealing.AnnealingSchedule	exponentialSchedule(int iterations, double startT, double endT)
NetworkState	getMaxAPosteriori(BayesianNetwork bayesNet, int sampleCount)
NetworkState	getMaxAPosteriori(BayesianNetwork bayesNet, int sampleCount, SimulatedAnnealing.AnnealingSchedule annealingSchedule) Finds the MAP using the default annealing schedule, which is an exponential decay schedule.
static SimulatedAnnealing	withDefaultConfig()
static SimulatedAnnealing	withDefaultConfig(KeanuRandom random)

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Detail

SimulatedAnnealing

public SimulatedAnnealing()

Method Detail

withDefaultConfig

public static SimulatedAnnealing withDefaultConfig()

withDefaultConfig

public static SimulatedAnnealing withDefaultConfig(KeanuRandom random)

getMaxAPosteriori

public NetworkState getMaxAPosteriori(BayesianNetwork bayesNet,
                                      int sampleCount)

getMaxAPosteriori

public NetworkState getMaxAPosteriori(BayesianNetwork bayesNet,
                                      int sampleCount,
                                      SimulatedAnnealing.AnnealingSchedule annealingSchedule)

Finds the MAP using the default annealing schedule, which is an exponential decay schedule.

Parameters:

bayesNet - a bayesian network containing latent vertices

sampleCount - the number of samples to take

annealingSchedule - the schedule to update T (temperature) as a function of sample number.

Returns:

the NetworkState that represents the Max A Posteriori

exponentialSchedule

public static SimulatedAnnealing.AnnealingSchedule exponentialSchedule(int iterations,
                                                                       double startT,
                                                                       double endT)

Parameters:

iterations - the number of iterations annealing over

startT - the value of T at iteration 0

endT - the value of T at the last iteration

Returns:

the annealing schedule